Optimizing Graph Neural Network With Multiaspect Hilbert-Schmidt Independence Criterion.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Mining

Background:

• Graph neural networks (GNNs) excel in data mining but struggle with integrating irrelevant neighbor node features.
• This feature integration leads to suboptimal node embeddings and hinders downstream applications.

Purpose of the Study:

• To propose GNN-MHSIC, a novel framework to optimize GNNs by mitigating the propagation of redundant information.
• To enhance node embedding quality and improve the performance of downstream tasks in GNNs.

Main Methods:

• Introduced the Hilbert-Schmidt independence criterion (HSIC), a non-parametric dependence measure, guided by the information bottleneck principle.
• Utilized HSIC to guide multi-aspect information propagation across GNN layers, minimizing input-layer dependence, maximizing layer-ground truth dependence, and minimizing layer-layer dependence.
• Theoretically proved finite upper and lower bounds for GNN-MHSIC.

Main Results:

• GNN-MHSIC effectively minimizes the propagation of redundant features while preserving essential information for target nodes.
• Experimental evaluations on four GNN models (GCN, GAT, HGAT, HGT) across three heterogeneous graphs demonstrated significant performance improvements.
• The framework proved beneficial across diverse GNN architectures and heterogeneous graph structures.

Conclusions:

• GNN-MHSIC offers a robust optimization strategy for GNNs, addressing the challenge of irrelevant feature propagation.
• The framework enhances the utility and performance of GNNs in various data mining and machine learning tasks.
• This research provides a theoretically grounded and experimentally validated method for improving GNNs.